Chain of thought
Chain of thought is the step-by-step reasoning a model writes out before giving an answer. Researchers read it to understand how a model reached a result and to spot unwanted behavior.
Also known as: chain-of-thought, reasoning trace
Many reasoning models work through a problem in intermediate steps before answering. That written reasoning is called the chain of thought.
Safety researchers use the chain of thought to monitor models. Techniques that make the reasoning shorter or unreadable can improve efficiency, but they also make that monitoring harder.